Introduction

In this notebook we train the dynamic part of the cross-domain model. The dynamic maps in the Carla dataset are the night, evening, and rainy routes. We train this by using a Siamese network and triplet loss. The non-grey parts of following image denote the dynamic network as a part of the bigger model. Trainig of the dynamic model occurs in a rotating fashion. (For the code please check networks/crossdomain_dynamic.py)

Dynamic Part

Setup

Setting paths

Loading image arrays and precalculated statistics

Creating the dataframe

Creating Dataloaders

Augmentations

Loading noon feature vectors

Load model

Load embeddings (uncomment when embeddings available)

Generate embeddings

Verification via PCA

If the interactive plot is not viewable please check train_dynamic.html in the html folder.

Save embeddings

Defining DataLoader

Plotting examples

Model and training loop

Load model (uncomment when a model is available)

Training and Saving

Saving

Proof of Concept (for rotating training procedure)